Perceiver: General Perception with Iterative Attention
Biological systems perceive the world by simultaneously processing high-dimensional inputs from modalities as diverse as vision, audition, touch, proprioception, etc. The perception models used in deep learning on the other hand are designed for individual modalities, often relying on domain-specific assumptions such as the local grid structures exploited by virtually all existing vision models.
Also cited · not yet reviewed (8)
- Transformer2017 · cited 7×, 3 in Method“Our latent Transformer uses the GPT-2 architecture (Radford et al. 2019), which itself is based on the decoder of the original Transformer architecture (Vaswani et al. 2017).”From this paper · §Methods
- ResNet2015 · cited 3×, 1 in Method“In contrast, the ConvNets that are typically used in image processing – such as residual networks (ResNets) (He et al. 2016) – bake in 2D spatial structure in several ways, including by using filters that look only at lo…”From this paper · §Methods
- Set Transformer2018 · cited 2×, 1 in Method“Attention is a permutation-invariant operation, and this property is preserved by the Perceiver and related models (Lee et al. 2019).”From this paper · §Methods
- Bahdanau attention2014 · cited 1×, 1 in Method“Both cross-attention and Transformer modules are structured around the use of query-key-value (QKV) attention (Graves et al. 2014; Weston et al. 2015; Bahdanau et al. 2015).”From this paper · §Methods
Show 4 more
- ConvS2S2017 · cited 1×, 1 in Method“The latent array itself is initialized using a learned position encoding (Gehring et al. 2017) (see Appendix Sec.”From this paper · §Methods
- ALBERT2019 · cited 1×, 1 in Method“We note that weight sharing has been used for similar goals in Transformers (Dehghani et al. 2019; Lan et al. 2020).”From this paper · §Methods
- ViT2020 · cited 4דThis is the strategy taken by the Vision Transformer (ViT) (Dosovitskiy et al. 2021), which first reduces the input size to ∼200\sim 200 using a 2D convolutional layer (referred to as “linear projection of flattened patc…”From this paper · §Related Work
- GoogLeNet (Inception)2014 · cited 2דImageNet has been a crucial bellwether in the development of architectures for image recognition (Krizhevsky et al. 2012; Simonyan & Zisserman 2015; Szegedy et al. 2015; He et al. 2016) and, until recently, it has been d…”From this paper · §Experiments
Led to
- Flamingo2022 · cited 2×, 1 in Method“Similar to Perceiver [48] and DETR [13], we learn a predefined number of latent input queries which are fed to a Transformer and cross-attend to the visual features.”From Flamingo · §Approach
Abstract
Biological systems perceive the world by simultaneously processing high-dimensional inputs from modalities as diverse as vision, audition, touch, proprioception, etc. The perception models used in deep learning on the other hand are designed for individual modalities, often relying on domain-specific assumptions such as the local grid structures exploited by virtually all existing vision models. These priors introduce helpful inductive biases, but also lock models to individual modalities. In this paper we introduce the Perceiver - a model that builds upon Transformers and hence makes few architectural assumptions about the relationship between its inputs, but that also scales to hundreds of thousands of inputs, like ConvNets. The model leverages an asymmetric attention mechanism to iteratively distill inputs into a tight latent bottleneck, allowing it to scale to handle very large inputs. We show that this architecture is competitive with or outperforms strong, specialized models on classification tasks across various modalities: images, point clouds, audio, video, and video+audio. The Perceiver obtains performance comparable to ResNet-50 and ViT on ImageNet without 2D convolutions by directly attending to 50,000 pixels. It is also competitive in all modalities in AudioSet.